Claude Skills are pluggable folder-based units that contain a SKILL.md file (with name and description in YAML frontmatter) plus optional scripts and reference files. Anthropic introduced Skills on October 16, 2025; Claude scans the metadata of every installed Skill and loads the body only when a request matches.
This guide is built off the May 11, 2026 JOEY video The TWO Claude Skills that run our entire AI video pipeline which hit 26,750 views on a 2,750-subscriber channel in three days (9.7x views-per-sub). The pattern matters because Skills moved from a developer-only feature to a marketer-shaped video primitive inside a single 7-day window.
Companion reading: the parent definitional hub Claude Code for Marketers covers the agent itself; the small-business guide Claude for Small Business Marketing covers Anthropic's May 13 SMB launch and the 15 Skills shipped in that bundle.
What a Skill Is (90 Seconds)
A Skill lives on disk as a folder. The folder has one required file (SKILL.md) and any number of optional supporting files. Here is the minimum shape, verbatim from the Anthropic Agent Skills documentation:
higgsfield-image-prompts/
├── SKILL.md (required: name + description + body)
├── EXAMPLES.md (worked prompts and their outputs)
├── STYLE-LIBRARY.md (brand look modifiers)
└── scripts/
└── validate.py (optional: deterministic checks)SKILL.md opens with YAML frontmatter. Two fields are required:
--- name: higgsfield-image-prompts description: Write Higgsfield image prompts using the 6-anchor structure. Use when the request mentions Higgsfield, Nano Banana, or image generation for a scene. --- # Higgsfield Image Prompts ## The 6 anchors 1. Subject ... 2. Camera ... ...
The name field is max 64 characters, lowercase letters numbers and hyphens only. The description field is max 1,024 characters and is the only piece Claude reads at startup to decide if the Skill is relevant. A vague description never triggers; a specific one with concrete trigger words does.
How Skills Load (Progressive Disclosure)
Anthropic uses the term progressive disclosure: Skills load in three stages so context windows stay small. You can install dozens of Skills without paying a context tax for the ones not used in a given conversation.
| Level | When loaded | Token cost | Content |
|---|---|---|---|
| Level 1: Metadata | Always loaded at startup | ~100 tokens per Skill | name and description from YAML frontmatter |
| Level 2: Instructions | When Skill is triggered | Under 5,000 tokens | SKILL.md body with workflows and guidance |
| Level 3: Resources | As needed | Effectively unlimited | Bundled scripts and reference files accessed via bash |
Source: Anthropic Agent Skills documentation (platform.claude.com/docs).
The 2 Production Patterns from the May 2026 Outliers
Both patterns wrap a prompt-engineering recipe for a visual-generation model into a Skill. They are the two Skills JOEY ships in the May 11, 2026 video; the pattern generalises to any image or video model with non-trivial prompt vocabulary.
Pattern 1: Image-Prompt-as-a-Skill
Package the prompt-engineering recipe for an image-generation model into a Skill so every Claude session writes prompts in the same shape.
JOEY built a Skill that feeds Nano Banana Pro image prompts inside Higgsfield for a sci-fi K-pop video pipeline. Same anchors every time: subject, camera angle, lighting, setting, style modifier.
- - name: higgsfield-image-prompts
- - description: Write Higgsfield image prompts using the 6-anchor structure. Use when the request mentions Higgsfield, Nano Banana, or image generation for a scene.
- - SKILL.md body: the 6 anchors with one-line definitions, 5 verbatim example prompts, banned language rules.
- - EXAMPLES.md: 10 prompts plus the outputs they produced (link or describe).
- - STYLE-LIBRARY.md: brand-specific look modifiers (cinematic 35mm, anime cel-shade, photoreal product).
Image prompts drift if every chat starts fresh. The Skill keeps the structure stable, the brand look consistent, and Claude stops inventing camera language.
Pattern 2: Video-Prompt-as-a-Skill
Package the per-shot prompt structure for a video-generation model into a Skill so Claude writes 5-second clip prompts in the same vocabulary every run.
JOEY built a second Skill for Seedance 2.0 prompts: per-shot camera moves, action verb per shot, transition language, audio cues. Combined with the image Skill, the pipeline reduces wasted generation credits.
- - name: seedance-video-prompts
- - description: Write Seedance 2.0 video prompts with per-shot camera moves and action verbs. Use when the request mentions Seedance, video clip generation, or shot-by-shot prompts.
- - SKILL.md body: per-shot template, the verb list, camera-move vocabulary, transition phrases, audio cue format.
- - SHOTS.md: 8 worked examples of 3-shot to 5-shot sequences.
- - CAMERA-MOVES.md: full glossary (push in, orbit, whip pan, parallax, dolly zoom).
Video models punish vague verbs. A Skill that supplies the camera vocabulary, the action verb library, and worked multi-shot sequences keeps every generation closer to first-take usable.
JOEY's framing: "these skills will not eliminate having to spend hundreds and thousands of credits, but your output generations will be significantly better than trying to have Claude memorize every prompt style." The Skill is the persistence layer for prompt expertise.
How a Marketer Builds a Skill in 30 Minutes
The recipe assumes you already have a working prompt pattern you run more than twice a week. The 30 minutes is for packaging it into a Skill, not for inventing the pattern.
- 1
Pick the workflow worth packaging
Choose one recurring video-prompt task you run more than twice a week. Common starting points: writing image-gen prompts for product shots, writing video-gen prompts for 5-second clips, converting a script into a shot list.
- 2
Create the folder and SKILL.md file
Create a folder named with lowercase hyphens (max 64 characters), for example higgsfield-image-prompts. Inside, create SKILL.md. Add YAML frontmatter with name and description; description must say what the Skill does and when Claude should use it.
- 3
Paste your existing prompt patterns into the body
Below the frontmatter, write the procedural knowledge as plain markdown: the 6 prompt anchors that work (subject, camera, lighting, setting, style modifier, action verb), 3 worked examples, and any banned-language rules. Keep the body under 5,000 tokens.
- 4
Drop in reference files for examples and edge cases
Inside the Skill folder, add EXAMPLES.md with 5 to 10 verbatim prompts and the outputs they produced. Add EDGE-CASES.md for tricky cases (multi-subject scenes, brand-mark visibility). Files are read by Claude only when the SKILL.md body references them, so they do not consume context until needed.
- 5
Install the Skill in Claude Code or claude.ai
Claude Code: drop the folder into ~/.claude/skills/ (personal) or .claude/skills/ (project). Claude.ai: zip the folder and upload via Settings > Features (Pro, Max, Team, or Enterprise plan with code execution enabled).
- 6
Test with a real task and iterate
Ask Claude to generate three prompts using the Skill. Compare outputs against your past hand-written prompts. Edit SKILL.md to fix the gaps. The first iteration delivers ~70% of the value; the next two refine the long tail.
Anthropic ships a meta-Skill called skill-creator that builds the folder and SKILL.md scaffolding from a plain-English description. Worth installing first; everything else compounds from there.
Tools That Pair With Skills
A Skill is just the prompt-engineering layer. The actual video gets generated somewhere else. The four pieces that pair well in a 2026 marketer stack:
Higgsfield (Nano Banana Pro for images, Seedance 2.0 for video), Runway, Luma. The Skill writes the prompts; the tool generates the asset.
Suno (music), ElevenLabs (voice), or the native audio tracks Seedance 2.0 ships. A separate Skill can package the audio-prompt vocabulary.
Adobe Premiere, CapCut, or DaVinci Resolve. The Skill can output an edit-list (in/out points, caption text, transitions) the editor follows.
Brand Brain: AI post generation in your brand voice, content review queue, and scheduling for LinkedIn, X, Threads, Instagram. The layer above the video. Social media management.
Brand Brain does not produce video clips; it produces the brand-on text post, the carousel caption, the X thread, and the scheduled publish for the finished video. The split is producing on one side (Claude plus Skills plus the generation tools the Skills target) and brand-consistent multi-platform publishing on the other.
Skills vs MCP vs Subagents
Three terms that get conflated. Each is a different layer of the stack; most production pipelines use all three.
| Layer | Purpose | Loaded when | Lives as | Best for |
|---|---|---|---|---|
| Claude Skills | Procedural knowledge: workflows, prompt patterns, reference docs. | On demand when a request matches the description. | A folder with SKILL.md plus optional scripts and reference files. | Repeat workflows where the same recipe gets used across many chats. |
| MCP (Model Context Protocol) | Tool access: connecting Claude to external systems and APIs. | When Claude calls a tool exposed by the MCP server. | A separate server process that speaks the MCP protocol. | Reaching out to image generators, databases, GitHub, design tools. |
| Subagents | Delegated execution: spawning a fresh agent context for a parallel task. | When the parent agent invokes a subagent for a side task. | A spawnable process inside Claude Code. | Independent parallel tasks the parent should not lose context to. |
Plan Availability and Pricing
Per the Anthropic Agent Skills documentation, custom Skills on claude.ai are available on Pro, Max, Team, and Enterprise plans with code execution enabled. Claude Code Skills are filesystem-based and work on any plan that includes Claude Code access. Pricing is set at claude.com/pricing and changes occasionally; verified figures as of May 15, 2026:
- Free: $0 per month. Does not support custom Skill uploads.
- Pro: $20 per month billed monthly (or $17 per month billed annually).
- Max: from $100 per month.
- Team: $25 per seat per month billed monthly, or $20 per seat per month billed annually.
- Enterprise: $20 per seat plus usage at API rates; scales with model and task.
Check claude.com/pricing for the current numbers per tier.
Common Pitfalls
Writing a 200-line SKILL.md the first day
Anthropic recommends under 5,000 tokens in the SKILL.md body. Push examples and reference material into separate files inside the folder; Claude loads them only when the body links to them.
Vague description field
The description is what Claude scans to decide whether to load the Skill. "Helps with marketing" never triggers. "Write Higgsfield image prompts with the 6-anchor structure" does.
No worked examples
Skills without 5 to 10 concrete examples behave like rules nobody follows. Examples teach the agent the shape of good output far better than a paragraph of guidance.
Skills with secret hardcoded credentials
Anthropic warns Skills run in a code execution environment with file system access. Never paste API keys or credentials inside a Skill file; the agent reads them and may surface them in outputs.
Forgetting Skills do not auto-sync across surfaces
A Skill uploaded to claude.ai is not on the API. A Skill in Claude Code is not on claude.ai. Manage installation per surface; treat the SKILL.md folder as the canonical source and copy it into each environment.